volRegimeMM — Status & roadmap

Where this strategy is today

ENGINEERING COMPLETE — the code is correct, uniform and shippable. That is a statement about engineering, not about profitability; the performance verdict is separate and sits below it.

Dimension	State
Correctness	5 / 5 invariants pass
Console output	15 tagged calls, 0 off-format
Side panel	20 rows, padded to the 3-column grid
Chart marks	reference-levels
Palette	8 / 8 canonical
GUI settings	20 exposed · 0 read at runtime · 0 untunable · 0 inert
Deployed to a live install	Not yet — audited and staged, not yet shipped

Performance verdict. Over three years on 7 markets it returns a median -10.5% against buy-and-hold's 49.6% — an edge of -66.4% — and is profitable on 1 of 7. On this window it is not a replacement for simply holding the asset, and nothing in these reports claims it is.

What it actually does

Spot accumulation conditioned on the realized-volatility regime, not on price direction.

Four volatility estimators are computed each cycle and averaged into one σ — Parkinson (high-low range), a return-based rolling standard deviation, an ATR-normalised estimate, and an EWMA with λ = 0.97. That average is then z-scored against its own rolling history, and every decision reads the z-score rather than the raw σ:

• Accumulate when volZ ≤ VR_BUY_Z (−0.5) — calm relative to this market's own norm — one rung of a VR_MAX_DCA (6) ladder, each rung VR_DCA_DECAY (0.8) × the last.
• Scale out when volZ ≥ VR_SELL_Z (1.2) and the position is at least VR_GAIN (1%) up, selling 30–80% depending on how extreme the spike is.
• Take profit separately at 2× the gain target, selling half.
• Emergency stop when the position is VR_MAX_DD (15%) under break-even — sells the entire position at market.
• Halt after 3 consecutive losses, or a daily loss beyond 2% of allocation.

Using a z-score rather than an absolute threshold is the design decision that makes it portable: 'calm' for DOGE and 'calm' for BTC are very different numbers, and the z-score puts both on one scale.

This is the maintained version of volHarvester, renamed at v2.0 because the original documented itself as variance-risk-premium capture, which it is not. Both files still ship and both are selectable in the GUI; measured against each other on 3 years × 7 markets at 15m they are behaviourally identical — same trades, same returns, same final cash.

The whole result is one trade per market

Over three years and 111,044 bars of 15m data the strategy takes roughly eleven positions per market. Of the 88 exits:

• 49 take-profits, mean +2.54%, summed +125%
• 8 emergency stops, mean −19.40%, worst −32.00%, summed −155%

That is one stop per market — two on SOL — and it is the entire difference between this strategy and a profitable one. The median win rate is 91% and the median profit factor is 0.15. Those two numbers together are the diagnosis: it wins nine times out of ten and loses anyway, because a +2.54% win has to happen 7.6 times to pay for one −19.40% loss and it gets about six.

Median fee drag is 0.3% of the account. This is not a cost problem — unlike most of this pack the trading is far too infrequent for fees to matter. It is a risk-shape problem and it sits entirely in the exit.

The stop overshoots its own limit. It triggers at −15% and realises a mean of −19.40%, worst −32.00%. The trigger is evaluated on a closed bar, the whole position is then sold at market on the next bar's open, and on a 15m bar in a fast market that gap is several points wide.

The one profitable market is DOGE (+3.5%, profit factor 3.41), and it is also the market with the lowest exposure (54% against 70–100% elsewhere).

Correction, 2026-09-15: this page previously stated "As with the other accumulate-and-scale strategies in this pack, there is no stop-loss sell. Every sell path requires a profit." That is false, and it was the first thing an operator would read when deciding how much to commit. The emergency stop sells the entire holding at market whenever the position is 15% under break-even. It fired zero times in the earlier 60-minute measurement, which is how the claim survived: the earlier pass described the sell paths it had watched fire rather than the paths in the file. At the timeframe this strategy is actually run at, it fires once per market and decides the result. A roadmap item asking for "a way out of a position that never becomes profitable" has been removed for the same reason — the way out exists, and it is what loses the money.

Correction, earlier: the header claimed five volatility estimators including Garman-Klass and Yang-Zhang. Neither exists — volEstimates has four terms and yzVol was declared in the persisted store and never computed. The header has been corrected; read the array, never the header.

Read from the shipped strategy source, not from its documentation.

The measured constraint

Time in market is 81.4%, over 88 round trips across all 7 markets and three years. Everything in the performance report rests on that fraction of the bars.

Settings, measured

This strategy does not use the engine config surface.

What to build next

Ordered by expected effect on the result, not by effort.

1. The exit is the strategy — fix the stop before touching the signal

Highest priority

Median return is −10.5% over three years against a buy-and-hold of +49.6%, an edge of −66.4%. The median win rate is 91% and the median profit factor is 0.15.

Eight trades out of 88 produce that. Take-profits realise +2.54% on average across 49 exits; the eight emergency stops realise −19.40% on average, worst −32.00%. Summed, the wins are +125% and the stops are −155%.

So the shape to change is the ratio, and there are two independent ways in:

Stop overshoot. The stop is set at VR_MAX_DD = 15% and realises −19.40%. It is checked on a closed bar and then dumps the whole position at market on the next open. A staged exit — half at the trigger, the rest on the following bar — or a stop that also rests as a limit order would recover several points of that gap without changing when the decision is made.

Win size. VR_GAIN is 1% and the take-profit path fires at 2× that. A +2.54% mean win against a −19.40% mean loss needs an 88.4% win rate merely to break even, and the strategy achieves 91% — which is why it sits close to water rather than being catastrophic. Raising VR_GAIN trades win rate for win size; measure the pair together, never either alone.

Median fee drag is 0.3%, so unlike most of this pack there is no cost argument for trading less. The trading is already rare.

2. Retire volHarvester rather than shipping both

Highest priority

volHarvester is this strategy under its pre-v2.0 name. The two produce identical trades and identical returns on all seven markets at 15m, as they did at 60m, and both are separately selectable in the Gunbot GUI — so an operator can pick the retired one without knowing it is retired, and will then be reading a page that was written about a strategy documenting itself as something it is not.

The rename corrected documentation, not behaviour. There is nothing to migrate; there is only a file to stop shipping.

3. Measure exposure against outcome — DOGE is the tell

High priority

DOGE is the only profitable market (+3.5%, profit factor 3.41) and also the one with the lowest exposure at 54%, against 70–100% everywhere else. ETH sits at 100% exposure and loses 10.5%.

That is one market and not a result, but it is the cheapest hypothesis available and it is directly testable. The accumulate gate is volZ ≤ −0.5 with no ceiling on how much of the account the ladder may commit, so in a market whose volatility stays below its own norm for a long stretch the strategy ends up fully invested and stays there — which is exactly the state in which a 15% drawdown becomes a 15% account loss.

A maxExposure cap of the kind already applied to the Directional engine in this pack would bound it, and the pack has the tool. Run the ladder at 60%, 80% and no cap across all seven markets and read the profit factor, not the return.

4. Re-tune the ladder, which has never been tuned

High priority

VR_MAX_DCA (6) and VR_DCA_DECAY (0.8) were dead until this pack's ladder repair — a bare return above the accumulate path made it unreachable while a position was held, and this strategy holds one on most bars, so the ladder never advanced past its first rung and DCA_DECAY ^ 0 was always 1.

They work now, which means their shipped values have never been measured doing anything. Six rungs decaying at 0.8 puts 26% of the ladder's total into the first rung and 7% into the last; whether that distribution suits a strategy whose losses come from being too far in is an open question with a cheap answer.

5. Check whether four estimators beat one

Worth testing

The ensemble averages Parkinson, return-standard-deviation, an ATR-normalised estimate and an EWMA, then z-scores the average. Four estimators of the same quantity on the same bars are highly correlated, and averaging them mostly removes the differences that would make any one of them informative.

The experiment is one line: run the strategy on each estimator alone and on the ensemble, same seven markets, same window, and compare. If the ensemble does not beat its best single member, three quarters of the computation is decoration — and the header has already been wrong once about what is in this array.

6. Measure it against cash, not only buy-and-hold

Worth testing

The edge column compares this strategy to holding the asset, and over this window holding was strongly positive on six of seven markets, so the −66.4% edge is partly a statement about the window.

But the absolute number is the one that matters to an operator here: −10.5% median against cash, on a strategy that is in the market 81% of the time. It underperforms doing nothing at all, which no choice of benchmark rescues. Both figures belong on the page and the report prints both.

7. The settings measurement does not yet cover this lineage

Read before acting on the others

the presentation audit measures which settings a strategy actually reads by wrapping the shared settings accessor with a recording proxy. The 348 Gunbot Quant files route through the shared settings accessor; the Quantroduction files — this one included — build a plain P: { ... } from the GUI settings directly, so nothing is measured for them and the untunable and inert counts on this page are blank rather than zero.

Every dead setting found in this lineage so far was found by hand. Until the proxy covers P the same way, treat a blank knob section here as unmeasured, not clean — the distinction that has already hidden a defect class in this pack for months.